Holistic Continual Learning under Concept Drift with Adaptive Memory Realignment

TL;DR

The paper addresses continual learning under concept drift by exposing models to evolving distributions where previously learned representations can become obsolete. It introduces Adaptive Memory Realignment (AMR), a drift-aware memory-update mechanism coupled with a drift-detection module (based on $D_{\text{KS}}$ in the main setup) to selectively refresh memory for drifted classes. The authors provide a formal problem formulation, theoretical results on gradient misalignment and the limitations of reservoir sampling, and extensive experiments across drift-augmented benchmarks showing AMR achieves full-retraining-like accuracy with far less labeling and computational cost. The work contributes a scalable, architecture-agnostic solution that unifies memory-based continual learning with concept-drift adaptation, with practical impact for non-stationary deployment scenarios.

Abstract

Traditional continual learning methods prioritize knowledge retention and focus primarily on mitigating catastrophic forgetting, implicitly assuming that the data distribution of previously learned tasks remains static. This overlooks the dynamic nature of real-world data streams, where concept drift permanently alters previously seen data and demands both stability and rapid adaptation. We introduce a holistic framework for continual learning under concept drift that simulates realistic scenarios by evolving task distributions. As a baseline, we consider Full Relearning (FR), in which the model is retrained from scratch on newly labeled samples from the drifted distribution. While effective, this approach incurs substantial annotation and computational overhead. To address these limitations, we propose Adaptive Memory Realignment (AMR), a lightweight alternative that equips rehearsal-based learners with a drift-aware adaptation mechanism. AMR selectively removes outdated samples of drifted classes from the replay buffer and repopulates it with a small number of up-to-date instances, effectively realigning memory with the new distribution. This targeted resampling matches the performance of FR while reducing the need for labeled data and computation by orders of magnitude. To enable reproducible evaluation, we introduce four concept drift variants of standard vision benchmarks, where previously seen classes reappear with shifted representations. Comprehensive experiments on these datasets using several rehearsal-based baselines show that AMR consistently counters concept drift, maintaining high accuracy with minimal overhead. These results position AMR as a scalable solution that reconciles stability and plasticity in non-stationary continual learning environments. Full implementation of our framework and benchmark datasets is available at: github.com/AlifAshrafee/CL-Under-Concept-Drift.

Figures (16)

• Figure 1: Visualization of concept drift in continual learning. (a) Initial Distribution: The learning process begins with a class of kittens. (b) Incremental Task: A new task introduces adult dogs, prompting the model to form a decision boundary that separates kittens from dogs. (c) Concept Drift: Over time, kittens evolve into adult cats, and adult dogs are replaced by puppies. Although the class labels remain the same (cats vs. dogs), their visual representation shifts, requiring an update in the decision boundary to maintain correct classification.
• Figure 2: Flow of our proposed Concept-Drift Adaptive Memory Realignment method for continual learning under concept drift. The approach integrates an uncertainty-based drift detection module with adaptive memory management to selectively retain and update buffer samples in the presence of recurring classes exhibiting distributional shifts.
• Figure 3: Evolution of the task-1 feature space (two classes) across five tasks on CIFAR-10. Without drift (top) the classes remain linearly separable; with drift introduced at task 3 (bottom) the features collapse.
• Figure 4: Comparison of computational cost and accuracy for different drift adaptation strategies. AMR achieves near-equivalent accuracy to FR with significantly lower sample requirement, relative time and GFLOP consumption.
• Figure 5: Comparison of class-incremental accuracy across different experience replay methods and varying number of drifts using No Adaptation, AMR, and Full Relearning strategies. AMR consistently achieves comparable accuracy to FR while using significantly fewer labeled samples.